Overview
What the API is, what it is compatible with, and what it deliberately is not.
Cognitivers is an inference API for open-weight models served without a refusal layer. You send OpenAI chat completion requests to a different host and get answers from models that were not trained or patched to decline the task.
The shape of the API
| Base URL | https://api.cognitivers.com/v1 |
| Long-context streaming | https://stream.cognitivers.com/v1 |
| Authentication | Authorization: Bearer sk-cog-... |
| Compatibility | OpenAI chat completions, embeddings, server-sent events |
| Models | cog-fast, cog-pro, cog-embed |
| Health | https://api.cognitivers.com/health |
The same key works on both hosts. Use the streaming host for requests with very large inputs, where the model can spend a while reading before it writes the first token; the connection is held open with keep-alive signals so a long read is not mistaken for a dead one.
What works unchanged
Anything that lets you set a base URL. That includes the OpenAI Python and Node SDKs, the Vercel AI SDK, LangChain, LiteLLM, Aider, opencode, Goose, Crush, Pi, Oh my Pi, DeepSeek Harness, Codex, Zed, Continue, Cline, Roo Code and GitHub Copilot in VS Code. Each one has a page with the exact configuration in Tools.
Two cases are worth knowing before you start:
- Claude Code speaks the Anthropic Messages API rather than OpenAI chat completions, so it needs a local bridge. Its page explains the bridge and the one setting that keeps long streams alive.
- Cursor does not accept a third-party endpoint at all. Its page says so plainly instead of offering a workaround we cannot support.
What it is not
- Not a chat product. There is a console, but the API is the product.
- Not a moderation service. Nothing scores or filters your prompts. What you do with the output is governed by the acceptable use policy.
- Not unlimited. Context, concurrency and spend have per-plan ceilings, all of them in Plans.